Instruction tuning trains a model on examples of instructions and useful responses so it becomes better at following user requests.
Instruction tuning is additional training that teaches a model how to respond to instructions.
It usually starts from an already pre-trained model. The training examples pair a request with a desired response, such as “Summarize this paragraph.” followed by a short useful summary, or “Translate this into French.” followed by a correct translation.
Think of a knowledgeable employee learning how to follow workplace requests. Pre-training gives the employee broad knowledge and language; instruction tuning gives practice with “Summarize this,” “Classify that,” and “Answer in this format” until the employee follows requests more reliably.
Instruction: Summarize this article.
Response: A short summary.
Instruction: Classify this support ticket.
Response: Billing.
Instruction: Return JSON with name and date.
Response: Structured output.
Pre-training: Broad initial learning.
Instruction tuning: Additional training focused on following instructions.
Instruction tuning: Changes model parameters.
Prompting: Changes the instructions supplied at runtime.
Instruction tuning: Changes model behavior through training.
RAG: Supplies external context during use.
Explore next
Instruction tuning uses training examples to make a model better at following instructions.
A video explainer for Instruction Tuning is not available yet.